bbernhard
commented
5 years ago Thanks a lot for your suggestions, very much appreciated!
You are right, there are currently no component labels available. The last few days/weeks there was a bit less activity on the site, so I decided to use the time to do a bit of much needed code refactoring and maintenance stuff.
------ on a side note ---- I've added the possibilty to create image collections. Image Collections are a simple way to structure/group image uploads. In case you want to create a new Image Collection, navigate to your profile -> Collections.
After you've create a image collection, you can pass the image collection's name as parameter to the HTTP request to automatically add the image to your collection. (I'll later extend the documentation here with an example)
edit: I've also added your ./low-res-batch contributions to an image collection....so you should see an image collection in your profile.
regarding the component labels: I think there is no technical restriction why annotating a component label won't work. So I guess, what I could do is, make a few component labels (like wheel/car paw/dog, ..) productive. That way, one can already start working on those labels, while we incrementally improve the visual appearance (label hierarchy).
I also like the idea of showing cropped images in case one annotates a component...I think that would be pretty cool. Still have to think, whether the cropping should be done on the server side or the client side...
dobkeratops
some general comments & ideas ...
(i) r.e the importance of component labels:- (a) could they accelerate training: because they narrow down feature detectors more? (b) could they give more expressive power with the curated label list, i.e. labelling components even if you dont have the labels for the object yet ("wheel / container" (.. on a wheelie-bin , or on a wheelbarrow.)
(ii) regarding the current workflow I still think components are hard to use i.e. i haven't been able to find them in search mode; it also doesnt' work well to automatically add components for the main label (because the system doesn't know which components are visible)
(iii) I still think Label-Me's hierarchy idea is extremely useful (although this is a whole extra dimension to UI, so it might be a lot of work): it lets you ground the annotations into a tree - this could sidestep the need for seperate "head/dog", "head/person" etc labels - instead what you do is label a "head", (and "paw", "tail" ..) ,and group it under an enclosing "dog" label. This could work well re. (i)(b). e.g. if you have a "wheel", "handle", "container" labels, then group them as one unit, you've described a "wheelbarrow", without needing the 'wheelbarrow' label yet
regarding UI there's the complexity of code but also user discovery. LabelMe uses drag-drop in a tree view, I think that's fairly intuitive .. but it still isn't obvious. I was trying to do it with lines to indicate parent-child links (more of a general purpose geometry editor) and I dont think would be too complex for casual users.
There's another idea which might suit your app better, which is to present a cropped/labelled region as a new image, then labels added within it could be considered children; perhaps you could say "annotate parts of this dog" to make it clear you shouldn't label parts behind/infront of it in this view.
failing anything like that, just being able to annotate components without grouping would still be useful IMO.
(iv) some suggestions for more component labels:-